METHOD AND SYSTEM FOR CREATING A PREDICTIVE MODEL FOR TARGETING WEB-PAGE TO A SURFER
Systems and methods for determining predictive model types are provided. A method may include generating a predictive model for a web page of a web site, wherein the web page includes a configuration defining one or more objects presented with the web page, and wherein each object is associated with a predictive model. The method may include determining one or more predictive model types that are associated with the predictive model, determining one or more performance indicators that correspond to each determined predictive model type, wherein performance indicators represent one or more benefits to a website, selecting a predictive model type of the predictive model out of the one or more predictive model types, wherein the predictive model type is selected based on a performance indicator corresponding to the selected predictive model type, and determining a configuration of the web page using the selected predictive model type of the predictive model.
Latest LivePerson, Inc. Patents:
This application is a continuation of U.S. patent application Ser. No. 14/753,496, filed on Jun. 29, 2015, which is a continuation of U.S. patent application Ser. No. 14/275,698, filed on May 12, 2014, now U.S. Pat. No. 9,104,970, which is a continuation of U.S. application Ser. No. 13/157,936, filed on Jun. 10, 2011, now U.S. Pat. No. 8,762,313, which is a continuation-in-part of U.S. patent application Ser. No. 12/504,265, filed on Jul. 16, 2009, now U.S. Pat. No. 8,799,200, which claims the benefit of and priority to U.S. Provisional Application No. 61/083,551 filed on Jul. 25, 2008 and U.S. Provisional Application No. 61/083,558 filed Jul. 25, 2008. This application is also a continuation-in-part of U.S. application Ser. No. 15/091,018, filed on Apr. 5, 2016, which is a continuation of U.S. patent application Ser. No. 14/313,511, filed on Jun. 24, 2014, now U.S. Pat. No. 9,336,487, which is a continuation of U.S. patent application Ser. No. 12/504,265, filed on Jul. 16, 2009, now U.S. Pat. No. 8,799,200, which claims the benefit of and priority to U.S. Provisional Application No. 61/083,551 filed on Jul. 25, 2008 and U.S. Provisional Application No. 61/083,558 filed Jul. 25, 2008. This application is also a continuation-in-part of U.S. patent application Ser. No. 14/582,550, filed on Dec. 24, 2014, which is a continuation of U.S. patent application Ser. No. 13/563,708, filed on Jul. 31, 2012, now U.S. Pat. No. 8,954,539, which is a continuation of U.S. patent application Ser. No. 12/503,925, filed on Jul. 16, 2009, now U.S. Pat. No. 8,260,846, which claims the benefit of and priority to U.S. Provisional Application No. 61/083,551 filed on Jul. 25, 2008 and U.S. Provisional Application No. 61/083,558 filed Jul. 25, 2008. All of these applications are hereby incorporated by reference in their entireties for all purposes.
BACKGROUND OF THE INVENTIONMore and more people are communicating via the Internet and other networks. The Internet, in particular, is a hierarchy of many smaller computer networks, all of which are interconnected by various types of server computers. Some of the servers interconnected through the Internet provide database housing as well as storage of a plurality of web-pages, generally referred to as the World Wide Web (WWW). By virtue of being accessible through the WWW, these web-pages may be retrieved by random Internet users, i.e. surfers, operating computers with a browser. Exemplary computers can be a personal computer (PC), a laptop, a notebook computer, a cellular telephone, a handheld computer, a personal data assistant (PDA), or any other computing device with wired or wireless communication capabilities communicable over an IP network.
Some common examples of browser applications used by Internet surfers are Openwave Systems Inc. or Opera Mobile Browser (a trademark of Opera Software ASA), Microsoft Internet Explorer (a trademark of Microsoft), and Firefox Web Browser. Using a web browser application on a computer that is connected to the Internet, surfers may retrieve web-pages that include information such as news, professional information, advertisements, e-commerce links, content downloads, etc. A common browser application may use Hyper Text Transport Protocol (HTTP) in order to request a web-page from a website. Upon receipt of a web-page request by a browser, the website responds by transmitting a markup language file, such as Hypertext Markup Language (HTML), that is representative of the requested page or pages. Notably, HTTP requests and HTML responses are well known in the art and are used as representative terms for the respective markup language files throughout this disclosure.
A common web-page may include numerous buttons, or links, operable to redirect a surfer to other locations within the website or on the Internet. These links offer a surfer a path to a next event which may be the presentation of another web-page, embedded content within the present web-page (e.g. an image, an advertisement, a banner, etc.), a playable media file, a number for forwarding a short message service (SMS) with a password, an application form, a registration form, etc.
A common link design may display a name of a category of information, such as news, sport, or economics. Other link designs may comprise a banner or an image intended to attract the attention of a surfer to an advertisement, an image, a text message that prompts the surfer to dial a number and use a password, etc. If a surfer is enticed to explore the offer shown on the link design, the surfer will use a pointing device such as a mouse, and place the pointer over the selected button, which may be comprised of a banner for example, and issue a command by “clicking” the mouse to “click through” on that button. In such a scenario, the surfer's browser may return information from a website associated with the particular banner that comprised the link.
In the present description, the terms “selecting button,” “selection button,” “redirecting button,” “slot”, “link,” “Hyper Link” and “banner” are used interchangeably. The terms “banner” and “slot” can be used as a representative term for the above group. An “advertisement” (AD) or “object” may be used as representative terms for content. Exemplary types of content can be the text of an AD as well as an AD's font, color, design of the object, an image, the design of the page in which the object is presented, etc.
The benefit from presenting a web-page, as well as improving a surfer's experience when surfing the web-page, can be increased if the selection buttons within the web-page are targeted toward the immediate observer. There are numerous existing methods and systems for offering targeted content in a web based environment. Some of the methods employ questionnaires containing categorized questions on user preferences. Such methods require the management of huge databases containing information on a large number of users. Besides the cumbersome management of all the information acquired from questionnaires, another negative is that many users prefer not to reply to a personalize questionnaire.
Other methods for identifying and offering targeted content in a web environment make use of client applications installed on a user's computer. The client applications are operable to track a user's activity on the web and subsequently report a compilation of the tracked activity to an associated web server or content server. Such methods are not popular with many users concerned with privacy. Further, such methods may require the often costly and awkward installation of a client application on a user's computer.
Some methodologies for delivery of targeted content may comprise a learning period and an ongoing period. During the learning period, a plurality of options of content within a certain web-page may be presented to various surfers. The response of the surfers to the various content options is monitored throughout the learning period in anticipation of ultimately employing the best performing alternative. During the subsequent ongoing period, all surfers requesting the certain web-page will be exposed to the previously selected alternative. Such an algorithm is often referred to as “The king of the Hill” algorithm. While a “King of the Hill” approach can fit the preference of a large group of surfers, it is prone to missing other groups of surfers that prefer other content delivery alternatives.
BRIEF SUMMARY OF THE INVENTIONThere is a need in the art for a method and system that can calculate predictive models for determining an object to be presented in association with a given link in a web-page used to attract an observer and entice a response. Exemplary responses may be the selection, or clicking of, a certain link (banner, for example), sending of an SMS, calling a number and using a password, the amount of revenue that was created by selecting this link, the number of visits of a surfer in a certain page, which were needed until the revenue was created, converting rate, etc. In an embodiment of the present disclosure, when calculating the predictive model the expected value to the content provider is taken into account. In order to save expenses and complexity, it is desired that the system be transparent to a surfer and will not require a database for storing information associated with a number of surfers. Converting is the state in which a surfer converts a link according to the desire of the website owner. Purchasing is an example for converting, filling a questionnaire can be another type of converting, registration, is another type of converting and so on.
Exemplary embodiments of the present disclosure seek to provide novel solutions for determining which content object taken from a group of content objects will be best suited for presentation in association with a link on a requested web-page. Exemplary types of content objects may comprise the text, topic, font, color, image, or other attribute of an advertisement. Still other content objects may comprise the specific design of the object, an image, the design of the page in which the object is presented, etc. Other embodiments may determine how to respond to a surfer request from the website. The decision process for selecting a content object can be based on information that is associated with the request, such as a common HTTP request. Exemplary associated information may include the day and time of receipt of the request for the web-page, the IP address and\or domain from which the request was sent, the type and the version of the browser application that is used for requesting the web-page, and the URL used by a surfer for requesting the web-page with the parameters that are attached to the URL.
Other types of information that is associated with the request can be statistical information indicative of the behavior of a surfer in relation to the website to which the request was sent. Behavioral information may include timers, frequency of visits by a surfer to the website, the last time that a surfer requested a web-page, etc. Other behavioral information may include one or more counters wherein each counter can count the number of events of a certain type. Exemplary counters can measure the number of visits by a specific surfer to the relevant website, the number of requests for a certain web-page within the website, the number of times a certain offer (content object) was selected or not selected, etc. Even further, some of the exemplary counters may be time dependent such that the value of the counter descends over time. In some exemplary embodiments of the present disclosure, the behavioral information may be embedded within a cookie that is related to the relevant website or a third party cookie.
Still other exemplary types of information comprise grouping or categorizing information. Grouping information can be delivered by a web server that contains a requested web-page. The grouping information can be associated with a group to which a current surfer belongs, surfer's grouping information (SGI). Exemplary surfer's grouping information can be used to define attributes of the group such as preferred sport clothing, preferred brand names, marital status, gender, age, etc. Surfer's grouping information can be managed by the web server and/or by a 3rd party server, and added to a field in a cookie associated with a certain surfer, for example.
Other grouping information can reflect attributes of the content, content grouping information (CGI). The CGI can be related to the requested web-page as well as attributes of the content objects presented in the requested web-page. Exemplary CGI can define attributes of the page or the content object such as the cost of a product, a product brand, vacation information, sport information, etc. CGI can be managed by the web server and added to a field in the URL of a certain web-page or URL of a certain content object, for example.
Each type of information, associated information, statistical information and grouping information can be retrieved and processed for defining one or more predictive factors which can be used in one or more predictive models. The predictive factors are used for calculating the predictive value gained by the website when presenting each content element. This value can also be the probability of a certain content object, from a group, to be selected, to be clicked, by a surfer currently observing the requested web-page. The predictive factors can be referred to as extracted features.
Some embodiments use a utility value or a key-performance indicator (KPI). A utility value, or the KPI value, can represent a website's benefit when the alternative is explored by a surfer. The benefit can be the rate of clicking (selecting) on a certain object in the presented web-page, the rate of converting a certain object, purchasing rate, the revenue rate which was created by presenting the web-page, etc. The probability can be multiplied by the associated utility to obtain the expected utility when presenting the alternative. In the present description, the terms “predictive factor,” and “predictive variable” may be used interchangeably.
In some embodiments the predictive model of a certain web-page takes into consideration the visit of a surfer in the website and the number of times the same web-page has been requested until creating a utility value, a certain KPI, related to the web-page. A visit of a surfer started from the first request to the website and ended after a certain period of the surfer's inactivity with the website. The period can be few tens of minutes of inactivity, 30 minutes for example. For a predictive model of a web-page that reflects the visit of the surfer in the website, the weight of the utility value or the KPI, which was created at the end of the visit, can be divided by the number of times the surfer observed the relevant web-page until determine to select, or purchase certain goods or services or create a certain revenue to the website owners.
An exemplary embodiment may build in parallel two or more predictive models for the same web-page. Each model can be related to a different KPI. Each model can be ready to be used in a different time. Building a predictive model per each KPI requires a certain level of variance within each KPI samples. In general the number of events in which the KPI is just clicking, fetching a certain object, is more frequent than the number of events in which a link is converted into a purchase or to deliver certain revenue. Therefore the required level of variance in the clicking samples can be reached in a shorter learning period than the learning period which is needed for preparing a predictive model for purchasing KPI. Consequently, the predictive model that is related to KPI which is just click on a link will be ready before a predictive model that is related to revenue KPI.
An exemplary embodiment that creates two or more predictive models per web-page, wherein each model is related to another KPI, can be able to determine when to switch from one predictive model to the other. The decision can be based on the readiness of the models and the utility of the relevant KPI to the web-page owners. For example, a predictive model that is related to revenue can be the last to be ready; however it can be the most important one to the website owner. Therefore, such an embodiment may start providing targeted content based on a predictive model that is related to “click” KPI and later on it may automatically switch to use a predictive model that is related to converting rate KPI and finally may use predictive model that maximize the revenue KPI, for example.
Revenue KPI can be measured as the amount of dollars, Euros, or other currency which was created from a certain web-page. In some embodiments the revenue KPI can be measured as the amount of dollars, Euros, or other currency which was created during a certain visit, etc.
An exemplary embodiment of the present disclosure can create a bank of predictive models. Each predictive model can be associated with a content object from the set of content objects that can be presented over the requested web-page. An exemplary predictive model may include one or more predictive factors with each predictive factor (predictive variable) being associated with a coefficient in a predictive formula. Exemplary predictive formulas can be based on known predictive algorithms such as, but not limited to, logistic regression, linear regression, decision tree analysis, etc. Some exemplary embodiments of the present disclosure can use linear or logistic regression, with or without stepwise methods, while calculating the predictive formula.
A predictive factor can also be a subset of values of certain variables such as, for example, the weekdays Monday and Saturday. The coefficient can thus outline the effect a predictive factor has on the predicted probability that a relevant content object will trigger a desired response from a surfer. The desire response, which is the KPI, can be the clicking, converting, creating revenue, etc. Exemplary predictive factors may include, for example, the day in the week (Monday and Saturday, for example), the hour, the browser type, the number of visits to the site, the content object that is presented in accordance with another link on the same web-page, the total elapsed time from the last visit, etc.
Exemplary predictive models can include some constants that are related to the content object associated with the model. Exemplary constants may be a utility constant which reflects the benefits the owner of the website receives when the relevant content object is selected or an arithmetic factor.
For each content object presented over a requested web-page, a predictive model with relevant predictive factors is processed such that the predicted objective, the probability of success for example, is calculated. A success is defined as a surfer responding positively to the presented content. For example, should a surfer select a relevant content object, the probability of the objects that can be presented is calculated. Subsequently, the combination of one or more objects with the highest predictive expected utility are selected to be associated with the links in the web-page requested by the surfer. The markup file that represents the web-page is modified such that its links point to the selected objects. The predictive objective value can be calculated to correspond to the predictive model. For example, in the logistic regression predictive model, the optimal linear predictive function is calculated and converted using a link function such as “Log Odds,”
An exemplary embodiment of the present disclosure may include a learning module. An exemplary learning module may be adapted to monitor the data exchange between a plurality of random surfers at one or more websites. Further, it can collect information on objects embedded and offered within the requested web-pages as well as track how each of the random surfers responds to those offers. From time to time, the exemplary learning module can process the sampled data in order to determine which predictive factors are relevant for each one of the objects and calculated an associated coefficient for success. Per each object, its associated statistical module can be updated or recalculated using the new or updated predictive factors.
An exemplary embodiment of the present disclosure can operate in either of two modes of operation, i.e. learning mode and ongoing mode. The learning mode can be executed after the initialization or when the content of the website is changed, for example. During the learning mode, new predictive models are calculated. The ongoing mode can be executed after the learning mode and may monitor and tune the existing predictive models.
During the learning mode, a large portion of communication sessions are sampled in order to define the new predictive models. During the ongoing mode, the size of the sample can be reduced and the predictive model can be tuned. When a significant change in the performance of a predictive model is observed, then a notification can be issued and/or a new predictive model can be created.
Some embodiments, during the learning mode, may collect information that is related to a visit. The information may include the number of times that the relevant webpage was retrieved by a certain surfer during a certain visit, etc. There are embodiments that during the learning mode collect information that related to different KPI. Such embodiments may deliver two or more predictive models per web-page. Each predictive model can be associated with different KPI and each predictive model can be ready in different time. Such embodiments are capable of switching between predictive models based on their readiness and their utility to the website owner.
Unless otherwise defined, all technical and/or scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the invention pertains. Although methods and materials similar or equivalent to those described herein can be used in the practice or testing of embodiments of the invention, exemplary methods and/or materials are described below. In case of conflict, the patent specification, including definitions, will control. In addition, the materials, methods, and examples are illustrative only and are not intended to be necessarily limiting. Moreover, the language used in this disclosure has been principally selected for readability and instructional purposes, and may not have been selected to delineate or circumscribe the inventive subject matter, resort to the claims being necessary to determine such inventive subject matter. Reference in the specification to “one embodiment” or to “an embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiments is included in at least one embodiment of the invention, and multiple references to “one embodiment” or “an embodiment” should not be understood as necessarily all referring to the same embodiment.
Implementation of the method and/or system of embodiments of the invention can involve performing or completing selected tasks manually, automatically, or a combination thereof. Moreover, according to actual instrumentation and equipment of embodiments of the method and/or system of the invention, several selected tasks could be implemented by hardware, by software or by firmware or by a combination thereof using an operating system. Software may be embodied on a computer readable medium such as a read/write hard disc, CDROM, Flash memory, ROM, etc. In order to execute a certain task, a software program may be loaded to an appropriate processor as needed.
The foregoing summary is not intended to summarize each potential embodiment or every aspect of the present disclosure. Other features and advantages of the present disclosure will become apparent upon reading the following detailed description of the embodiments with the accompanying drawings and appended claims.
Exemplary embodiments of the present disclosure will be understood and appreciated more fully from the following detailed description, taken in conjunction with the drawings in which:
In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the invention. It will be apparent, however, to one skilled in the art that the invention may be practiced without these specific details. In other instances, structure and devices are shown in block diagram form in order to avoid obscuring the invention. References to numbers without subscripts or suffixes are understood to reference all instance of subscripts and suffixes corresponding to the referenced number. Moreover, the language used in this disclosure has been principally selected for readability and instructional purposes, and may not have been selected to delineate or circumscribe the inventive subject matter, resort to the claims being necessary to determine such inventive subject matter. Reference in the specification to “one embodiment” or to “an embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiments is included in at least one embodiment of the invention, and multiple references to “one embodiment” or “an embodiment” should not be understood as necessarily all referring to the same embodiment.
Although some of the following description is written in terms that relate to software or firmware, embodiments may implement the features and functionality described herein in software, firmware, or hardware as desired, including any combination of software, firmware, and hardware. In the following description, the words “unit,” “element,” “module” and “logical module” may be used interchangeably. Anything designated as a unit or module may be a stand-alone unit or a specialized or integrated module. A unit or a module may be modular or have modular aspects allowing it to be easily removed and replaced with another similar unit or module. Each unit or module may be any one of, or any combination of, software, hardware, and/or firmware, ultimately resulting in one or more processors programmed to execute the functionality ascribed to the unit or module. Additionally, multiple modules of the same or different types may be implemented by a single processor. Software of a logical module may be embodied on a computer readable medium such as a read/write hard disc, CDROM, Flash memory, ROM, or other memory or storage, etc. In order to execute a certain task a software program may be loaded to an appropriate processor as needed.
In the description and claims of the present disclosure, “comprise,” “include,” “have,” and conjugates thereof are used to indicate that the object or objects of the verb are not necessarily a complete listing of members, components, elements, or parts of the subject or subjects of the verb.
The present disclosure relates to the presentation of data over communication networks, such as the Internet, and more particularly to selecting one or more objects (content of a web link) from a group of objects of content, to be presented over a requested web-page. An “object” may be used as representative terms for content. Exemplary types of content can be the text of an AD as well as an AD's font, color, design of the object, an image, the design of the page in which the object is presented, etc.
Turning now to the figures in which like numerals represent like elements throughout the several views, exemplary embodiments of the present disclosure are described. For convenience, only some elements of the same group may be labeled with numerals. The purpose of the drawings is to describe exemplary embodiments. Therefore, features shown in the figures are chosen for convenience and clarity of presentation only.
The communications system 100 may be based on the Internet Protocol (IP). Network 120 may represent one or more network segments including, but not limited to, one or more Internet segments, one or more Intranets, etc. Network 120 may run over one or more types of physical networks such as, but not limited to, Public Switched Telephone Network (PSTN), Integrated Services Digital Network (ISDN), cellular networks, satellite networks, etc. Further, network 120 may run over a combination of network types. Network 120 may also include intermediate nodes along the connection between a surfer and a content server. The intermediate nodes may include, but are not limited to, IP service provider servers, cellular service provider servers and other types of network equipment.
It will be appreciated by those skilled in the art that depending upon its configuration and the needs, communication system 100 may comprise more than three ST 110, three PCS 140 and three PCD 150. However, for purposes of simplicity of presentation, three units of each are depicted in the figures. Further, it should be noted that the terms: “surfer terminal”, “terminals,” “endpoint computer,” “endpoint,” “surfer,” “user's device,” “mobile phone” and “user” may be used interchangeably herein.
A plurality of ST 110 may be served by system 100 for surfing the Internet 120 and fetching web-pages from the one or more PCS 140 or PCD 150. Exemplary ST 110 can be a personal computer (PC), a laptop, a notebook computer, a cellular telephone, a handheld computer, a personal data assistant (PDA), or any other computing device with wired or wireless communication capabilities communicable over an IP network. A common ST 110 may run a browser application such as, but not limited to, Openwave Systems Inc. or Opera Mobile Browser (a trademark of Opera Software ASA), Microsoft Internet Explorer (a trademark of Microsoft), or Firefox Web Browser, etc. The browser application can be used for rendering web-pages. Exemplary web-pages may include information such as news, professional information, advertisements, e-commerce content, etc.
A common browser application can use HTTP while requesting a web-page from a website. The website can respond with a markup language file such as but not limited to HTML files. Herein the term HTML is used as a representative term for a markup language file. HTTP requests and HTML responses are well known in the art.
Exemplary PCS 140 and/or PCD 150 may receive HTTP requests from the plurality of ST 110 and deliver web-pages in the form of markup language files such as, but not limited to, HTML files. An exemplary PCS 140 may comprise a surfer's interface unit (SIU) 143 and a content adapting module (CAM) 146. Exemplary SIU 143 can execute common activities of a content server. Further, it may receive HTTP requests and respond with HTML files. In addition, exemplary SIU 143 may communicate with CAM 146 and/or deliver information about the activities of the different ST 110. The activity can be the web-pages requested by the surfers, for example.
In some exemplary embodiments of the present disclosure, SIU 143 may deliver, to CAM 146, information about surfer's attributes. Exemplary attributes may be a surfer's purchasing habits (expensive, not-expensive, brand name, etc.). A specific surfer's information can be managed by the web server and be added to a field in the cookie associated with the particular surfer.
In some embodiments, the information can also be related to the requested web-page as well as particular attributes of the content objects to be presented in the requested web-page. Exemplary information that may define attributes of the web-page or the content object include, but are not limited to, the cost of a product, the brand of a product, vacation information, etc. This type of information can be managed by the web server and added to a field in the URL of a certain web-page or URL of a certain content object, for example.
CAM 146 may process requests of different surfers in order to create a plurality of predictive modules, wherein each predictive module can be associated with an optional object that is presented over a requested web-page in one of the slots (redirection-button) from a set of optional slots. The predictive modules can be used to select a set of one or more objects that maximizes an expected benefit to the site owner. In the present description, the terms “expected benefit” and “expected utility” can be used interchangeably. The sets of slots, in which an optional object can be presented, and the selected optional object to be presented in each slot (Slot/Optional-object) can be defined within the configuration of the web-page. The configuration of the object/redirection-buttons with the optimal prediction to be selected for the requesting surfer is transferred to SIU 143. In response, SIU 143 may modify the HTML file to include those objects in the relevant links.
Some embodiments of CAM 146 may handle two or more predictive models per web page. Each predictive model can be associated with different KPI and each predictive model can be ready in different time. Such embodiments are capable of switching between predictive models based on their readiness and their utility to the website owner. In some embodiments, a predictive model can target a weighted mix of KPIs.
There are some embodiments of CAM 146 that may create predictive model per web-page taking into consideration the visit of a surfer. In such embodiment the weight of the KPI, which was achieved at the end of the visit and is related to the relevant web-page, can be divided by the number of times the relevant web-page had been requested and rendered to the surfer during the visit until a decision was made by the surfer.
Exemplary PCD 150 can include a plurality of common content servers 152 and one or more content adapting servers (CAS) 154. CAS 154 can be connected as an intermediate node between the plurality of ST 110 and the content servers 152. In one exemplary embodiment, CAS 154 may be configured as the default gateway of the PCD 150. In another exemplary embodiment, CAS 154 may physically reside between the PCD 150 and the network 120.
In yet another exemplary embodiment of the present disclosure, a redirector may be included and CAS 154 can be configured as a transparent proxy. In such an embodiment, CAS 154 may be transparent to both sides of the connection, to ST 110 and to the content servers 152. In an alternate exemplary embodiment of the present disclosure CAS 154 may be used as a non-transparent proxy. In such an embodiment, the ST 110 can be configured to access CAS 154 as their proxy, for example. In an alternate exemplary embodiment of CAS 154, not shown in the drawing, can communicate with a website by using a certain application program interface (API). Yet, in an alternate exemplary embodiment of CAS 154, not shown in the drawing, can be invoked by a link to a JavaScript (JS). More information on those embodiments is disclosed below.
CAS 154 can intercept the data traffic between the plurality of ST 110 and content servers 152. CAS 154 tracks the behavior of a plurality of surfers in order to create predictive models for content objects to be associated with requested web-pages. Subsequently, when operating in an ongoing mode, the predictive models can be used for determining which objects to assign to a web-page requested by a given surfer.
Toward PCD 150, CAS 154 can process requests of surfers for web-pages from one or more of the content servers 152, for retrieving a plurality of predictive factors to be used in a plurality of predictive modules, similar to the predictive modules that are used by exemplary CAM 146. In the other direction, CAS 154 can process the responses (the HTML files, for example) and determine, based on the predictive models, which object to present. The HTML file, which represents the web-page, is modified to include those objects in the links. More information on the operation of PCS 140 and CAS 154 is disclosed below in conjunction with
The following paragraphs are related to
In one exemplary embodiment, CAS 200 may intercept the communication between the plurality of ST 110 and the plurality of content servers 152 (
After creating the predictive models, CAS 200 can use the models in order to determine which combination of optional objects can be associated with certain redirecting buttons in a requested web-page such that the profit of the owner of the PCD 150 is maximized. This combination of optional objects is referred to as a predictive configuration of the web-page. Markup language files, such as HTML files, are transferred from content servers 152 toward ST 110 and can be modified by CAS 200, for example, according to the predicted configuration. The profit of the owner can be measured by KPI parameter such as but not limited to: the clicking rate on an object of a delivered web-page according to the predictive configuration; the converting rate, the rate of converting a link to a deal from the total number of events of presenting the page based on the predictive configuration; the revenue rate, the average revenue that is created by presenting the predictive configuration of a certain webpage compare to other configuration of the same web-page, etc. CAS 2000 of
The behavior of surfers can be monitored and learned during the ongoing operation of an exemplary CAS 200. The monitored data can be used in order to improve the predictive models.
An exemplary CAS 200 can be divided into two sections, a network interface section and a content adaptation section. The network interface section may comprise an HTTP proxy 210, a page request module (PRM) 220, an active surfer table (AST) 215, or 2150 for the example of CAS 2000, and a markup language file handler (MLFH) 230. The content adaptation section may include one or more page data collectors (PDC) 240, one or more page object selection modules (POSM) 250, and a management and prediction module (MPM) 260. CAS 2000 may include a PDC 2400 instead of PDC 240. PDC 2400 act in a similar way to PDC 240, however the collected data per web-page is organized according to the types of KPI. Further, CAS 2000 may include an AST 2150 instead of AST 215. Along the disclosure the term AST 215 may refer also to AST 2150 unless for certain cases it is written differently and the term PDC 240 may refer also to PDC 2400 unless for certain cases it is written differently. Data communication between the internal modules of CAS 200 may be implemented by using components such as, but not limited to, shared memory, buses, switches, and other components commonly known in the art that are not necessarily depicted in the drawings.
In an alternate exemplary embodiment of CAS 200 or 2000, not shown in the drawing, CAS 200 or 2000 can communicate with a website by using a certain application program interface (API). In such embodiment the surfer requests as well as the responded web pages are sent directly between the surfer and the website. The website, upon receiving a request parses the request, retrieved from the request information which is required by CAS 200 or 2000. Information such as the URL, the receiving time, the cookie, the browser type, etc. This information is sent from the website toward CAS 200 or 2000 by using the API. Then the website waits to get the instructions about the recommended configuration from CAS 200 or 2000. Upon receiving the information about the request, CAS 200 or 2000, process it, update the cookie, encode the cookie, determine which configuration is recommended and sent this data, by using the API to the website. The website in its turn, processes the received information from CAS 200 or 2000, defines the configuration of the webpage, and sends the webpage toward the requesting server.
Thus, such an exemplary embodiment of CAS 200 or 2000 may not need the HTTP proxy 210. Further, the PRM 220 and the MLFH 230 can be slightly modified to communicate with the website based on the API, processed the received information and responding to the website over the API.
Yet, in an alternate exemplary embodiment of CAS 200 or 2000, not shown in the drawing, CAS 200 or 2000 can be invoked by a link to a JavaScript (JS). In such embodiment the surfer requests as well as the responded web pages are sent directly between the surfer and the web-page. However the responded website includes one or more links. Each link requests a JS. The first link on the top of the webpage is a link to CAS 200 or 2000. In some embodiments the first link can point on a Cache Domain Network (CDN) server. The CDN server in its turn can transfer the request to the first JS to the CAS 200 or 2000. CAS 200 or 2000 can respond by sending a JS code that when it is activated by the browser at the surfer computer, is adapted to collect information that is required by the CAS 200 or 2000. Information such as the URL, the receiving time, the cookie, the browser type, etc. This information is sent by the browser running the JS from the surfer toward the CAS 200 or 2000. Upon receiving the information about the request, CAS 200 or 2000, processes it, determines a configuration of the requested web page, update the cookie, encode the cookie and send it to the JS. Updated the cookie can be done by adding information about the configuration that was sent to the user. In addition CAS 200 or 2000 determines which configuration to send to this surfer. In parallel the browser at the surfer continue parsing the web page and reaches the next link to a next JS, the browser send a request to the next JS, in response the CAS 200 or 2000 responds with the appropriate object to be presented.
Thus, the alternate exemplary embodiment of CAS 200 or 2000 the HTTP proxy 210, the PRM 220 and the MLFH 230 can be slightly modified to communicate with the surfer and process the requested JS.
An exemplary PCS 140 (
In exemplary embodiments of the present disclosure, in which the communication over network 120 is based on IP, HTTP proxy 210 may be adapted to handle the first layers of the OSI (open system interconnection) seven layer communication stack. The layers may be the Physical Layer, Data Link Layer, transport Layer (the TCP stack), and the Network Layer. In exemplary embodiments of the present disclosure, in which the CAS 200 is transparent to the ST 110 (
Data coming from the internal module of CAS 200 is transferred via HTTP proxy 210 toward their destination. For example, HTTP requests for web-pages, after being processed by PRM 220, are transferred toward the appropriate content server 152 (
Exemplary PRM 220 may comprise a bank of domain counters 222, a cookie decompression and update module (CDUM) 224 and a timer 226. PRM 220 may receive, from HTTP proxy 210, requests for web-pages that are targeted toward content servers 152 (
The collected information may include associated information such as, but not limited to, the day and the time the request for the web-page was received, the IP address or/and IP port from which the request was sent, the type and the version of the browser application used for requesting the web-page, and the URL used by a surfer for requesting the web-page with the parameters that are attached to the URL.
In addition, the retrieved information may include behavioral information. The behavioral information may be statistical information that is managed by CAS 200. It may be divided into a requester's related behavioral information and a domain's related information. The requester's related behavioral information may include timers indicating one or more previous visits by the requester to the domain PCD 150 (
The domain related behavioral information may include the number of requests for a certain web-page from the domain and the time of the last request for this page, the number of times a certain offer (content object) was selected, the last time that it was selected, etc. Some of the exemplary counting values may be time dependent such that the value of the counter may decrease over time. The domain related behavioral information may be counted and recorded in a plurality of domain counters 222 by PRM 220. In some exemplary embodiments of the present disclosure, behavioral information may be updated and written by CDUM 224 within a cookie that is related to PCD 150 (
The counters may be incremented by PRM 220 each time a request for an event that is related to the counter is identified by PRM 220. Incrementing the value of a counter may be time dependent such that the last value may decrease according to the time difference between the last event and the current event. Then the decreased value is incremented by one and the time of the current event is recorded. Fading the value of the counters may be done by using an exponential formula using a half-life-time constant, for example. The half-life-time constant may have a configurable value or an adaptive one. The half-life-time constant may be one property in a set of properties. It should be understood that the fading algorithm described above is offered herein for exemplary purposes only and should not be interpreted to limit the present scope. Those skilled in the art will appreciate that other exemplary embodiments of the present disclosure may use other fading algorithms for adjustment of a counter.
PRM 220 may retrieve an associated cookie from the request, if one exists. In one exemplary embodiment, in which the cookie is compressed, the cookie is transferred to CDUM 224. Those skilled in the art will appreciate that different methods may be used for compressing/decompressing information written in a cookie. An exemplary compression method may express the values of the counters using a subset of ASCII based characters in lieu of decimal based characters. Other embodiments may express the counters by using logarithmic numbers (integer and mantissa), for example. Still other embodiments may use the combination of the two.
The decompressed cookie is parsed by CDUM 224 to identify a requester's ID wherein a requester's ID is an ID number that has been allocated to the requester by PRM 220. The allocation of the ID can be done in the first request that was received from the relevant surfer in the current visit of the website. An exemplary ID may be defined randomly from a large number of alternatives. It may be a 32 bit, 64 or 128 bit number, for example. If a requester's ID exists in the cookie, then the AST 215 is searched for a section in AST 215 that is associated with the requester's ID. In some embodiments an ID can be allocated to a surfer at the start of the initial visit to the site. This ID can be kept in the cookie as long as the cookie exists. In such embodiment, a visit-ID can be added as a supplement to the surfer ID. If a section exists, then a new entry is allocated in the relevant section of AST 215 for recording the associated and behavioral information that is relevant to the current obtained requester's request. If a section in AST 215 was not found, or if an ID was not found in the cookie, a new ID may be allocated for the given requester. In some exemplary embodiments of the present disclosure, a field in the cookie may point to a file stored at the server that includes behavioral information associated with a surfer.
If a new requester's ID is allocated to the requester, then a new section in AST 215 may be allocated by PRM 210 to be associated with the new requester's ID and a new entry in the section is allocated for storage of the information that is related to the current request for a web-page.
Timer 226 may be used while managing behavioral information of a user. It may be used for timing indication on previous activities such as previous visiting, previous purchasing, etc., which were recorded in the cookie. Further, the value of timer 226 may be used in the process of time fading of the counters. These time records can be used as predictive factors. The clock of timer 226 may range from a few seconds to a few minutes, for example. Timer 226 may be a cyclic counter having a cycle of six months, for example.
Exemplary AST 215 may be a table, which is stored in a random access memory (RAM) that can be shared by the internal modules of CAS 200, for example. AST 215 may be divided into a plurality of sections such that each section is associated with a requester's ID and represents an active surfer. Further, each section may include a plurality of entries. Each entry may be associated with a request of a surfer from the website PCD 150 (
Each entry may include fields such as, but not limited to, the receipt time of the relevant request, associated information retrieved from the relevant request, etc. Exemplary associated information may be such as, but not limited to, the type and the version of the browser application used for requesting the web-page or the URL used by a surfer for requesting the web-page along with the parameters that are attached to the URL.
In some embodiments of the present disclosure, some the URLs may include content grouping information such as, but not limited to, attributes of the content objects. Exemplary attributes may be the cost of a product, the brand of a product, vacation information, etc.
Other fields in an entry of AST 215 may store updated requester related behavioral information. Still other fields may store domain related behavioral information that was valid when the previous web-page was sent from the domain to the requester. Yet another group of fields may store management information that is related to the operation of CAS 200 and indicate whether the current section is associated with a random surfer or a personalized surfer, for example. A common surfer may be a surfer for whom the object was presented as a result of using the object's prediction module. A personalized surfer can be referred also as common surfer. A random surfer may be a surfer for whom the object was selected randomly to be presented. A random surfer can be referred also as a control surfer. In some embodiments a random surfer can be defined per a slot, or one or more slots, in a certain web-page or in one or more web-pages for a current visit.
Some embodiments of AST 215 the entries may comprise also fields for storing information related to a current visit. Such field can store the starting time of the visit, the number of times that the URL associated with the entry was requested during a certain visit, etc.
In some embodiments of CAS 2000 the AST 2150 may have additional fields. These additional fields can be used for storing data that is related to different KPI. In such embodiments a feedback from the content server or a managing server of the domain may be received. The feedback information may include information regarding a purchasing decision, which was made in association with the URL of the entry. In some cases also the created revenue as well as an identification of the order of the purchase is delivered with the feedback and is stored at the relevant field of the entry, etc.
New sections and new entries in AST 215 are allocated by PRM 210. Different modules of CAS 200 may read and write information from/to the AST 215, as disclosed below relative to the description of the other modules of CAS 200. Once every configurable period, AST 215 may be scanned by MPM 260 looking for one or more inactive sections of AST 215. Exemplary periods may be in a range from a few seconds to a few tens of minutes. An inactive section is a section that is associated with a visit of a surfer in the web site. The decision when the visit is terminated is when the time period between the last received request (the time associated with the last entry of the section) is longer than a certain configurable value. Entries of inactive sections of AST 215 may be released after the stored data in the entries of the visit (section) is copied to the appropriate PDC 240. The configurable value of the time period can be between tens of minutes to few hours, between 30 minutes to an hour for example. A common value is 30 minutes from the last activity in the site.
After allocating an entry for the current request in the appropriate section of AST 215, PRM 210 may write the retrieved associated information in the appropriate fields of the entry. A requester's related behavioral information may be retrieved from the decompressed cookie. The appropriate counters of the requester's retrieved related behavioral information may be updated. The received value of the appropriate counters may be manipulated to include the time affect of the period from the last visit (i.e., the fading affect). The time adapted value may be incremented by one, counting the current visit. The updated value may be stored in the relevant field of the entry. In some embodiment, in which the new entry includes a field for storing the number of times that this web-page has been requested during the new visit, the value one (reflecting the first request) is stored in that field.
PRM 210 may determine whether a requester that was assigned a new ID value can be designated as either a “control” or “common” requester. The decision may be based on configurable parameters, which may depend on the mode of operation of CAS 200, such as, for example, whether CAS 200 is operating in a “learning” mode or an “ongoing” mode. The decision on how to label the requester may be written in the appropriate field in the associated entry or section. This configurable parameter can be one of the parameters in a set of properties.
In the case that the requester has a valid ID, the URL associated with the request is parsed in order to determine which web-page was previously delivered to the requester. The determination of the previously delivered web-page may indicate the stimulus that prompted the user to make the current request. Once determined, the requester's section in AST 215 is searched for an entry that is associated with a previously delivered page (PDP). An entry in which the recommended combination includes a link to the current received request, for example. If such an entry is found, then an indication of “success” is written in association with the combination of the requested object and the link design (i.e., slot or redirection button). In other exemplary embodiment of a CAS, information on a PDP can be received from information stored in the cookie. Alternatively the information can be received from the web-site by using an API. The success indication may be marked in a response field that is associated with the combination object and slot of the entry that includes the PDP that prompted the current requested object. While processing the information that is stored in the AST, object and slot combinations that do not have a success indication may be marked as “failures.”
After processing the request, PRM 210 may replace the cookie, or write a new cookie if the request does not already include a cookie, with updated behavioral information. Subsequently, the request with the new cookie may be transferred to one of the content servers 152 (
An exemplary MLFH 230 may comprise a cookie update and compression module (CCM) 232. MLFH 230 receives, via HTTP proxy 210, markup language (ML) files such as HTML files, for example, that represent the requested web-pages sent from content servers 152 toward ST 110 (
For a “control” requester, based on the web-page ID, MLFH 230 may randomly configure the received ML file. The configuration may include the set of links (i.e., slots or redirection buttons) in the web-page and an assigned object for each one of the links in the set. The assigned object may be selected randomly from a group of objects that may be presented on the requested web-page, as determined by ID of the web-page. In the case that the requester is not a “control” requester, the location of the entry in AST 215 that is relevant to the current page is transferred to one of the POSM 250 associated with the page ID of the current received ML file. In response, POSM 250 may initiate a selection process for determining the predicted configuration (object and slot) of the received ML file. As an example, the predicted configuration may be the configuration which maximizes the predicted expected value obtained by presenting the particular configuration.
In some embodiments the expected value can be a KPI, which reflects the KPI of the relevant requested web-page taking into consideration the visit of a surfer. Yet in some embodiments of CAS 2000, POSM 250 may use different types of predictive models per different requested web-pages. The type of the predictive models can be based on the type of the KPI according to which the predictive model was built to maximize that KPI. In some embodiments of CAS 2000, in different period of time the POSM 250 may automatically determine which type of predictive model to use for the same requested web-page.
When MLFH 230 has defined a configuration (randomly selected or predicted), it may manipulate the received ML file in order to reflect the configuration. Per each set (object and slot), the URL of the selected object may be assigned to the relevant link (slot) in the ML file. In addition, the entry in AST 215 may be updated to include the configuration of the page. Fields of the relevant entry in AST 215 that need to be included in an updated cookie, such as, but not limited to requester ID, updated page ID counters, and timers, are retrieved from AST 215 by MLFH 230. Some of the cookie fields are compressed by CCM 232, according to one of the compression methodologies previously described. The compressed cookie is added to the header of the modified ML file. The modified ML file with the new cookie may then be transferred toward the requesting ST 110 (
Each PDC 240 is associated with a web-page that is stored in one of the content servers 152 (
In some embodiments of CAS 2000 the PDC 2450 may comprise two or more KPI-objects-historical DB (KPIOHDB) 2480a-n. Each of that two or more KPIOHDB 2480a-n can be used for storing data that is related to a different KPI. In such embodiments each KPIOHDB 2480a-n may include a plurality of database couples 249kas & kaf to 249kns & knf. Each couple is associated with an object (objects a-n that may be assigned to the page) and a type of KPI. Each couple of databases, one database, 249kaf for example, stores records of events in which a link to the relevant object appeared in a delivered web-page but the KPI related to the DB and the object was not achieved. The second DB 249kas stores records of events in which the relevant object was selected and the KPI related to the object and the DB was achieved, for example.
In some exemplary embodiments the number of DB per object can be other than two DB (success; fail). For example, when a KPI has several values, zero dollars, one to ten dollars, 10 to hundreds dollars etc. In such a case three or more DBs can be used, one per each range of revenue.
From time to time, one or more entries of an inactive section may be retrieved from AST 215 by MPM 260. These entries may be sorted according to the web-pages that are associated with the entry and copied to ELOG 242 of the relevant PDC 240. Wherein each PDC 240 is associated with one web-page of the one or more web-pages, each one of these entries, in AST 215, may be transferred as a record into ELOG 242 and the inactive entries released from AST 215. In some embodiments copying the inactive session can be implemented when it is determined that the visit which is associated with the section was terminated.
OMM 246 manages the stored records that are related to its associated web-page. In order to reduce the cost and the complexity of the PDC 240 storage volume, the storage volume may be divided into two types, short term (ELOG 242) and long term (the bank of databases OHDB 248). Periodically, once in a configurable transfer period (TP), an hour for example, the records in the ELOG 248 may be copied into appropriate ODB 249 and then released. The TP can be a parameter in a set of properties, for example.
In one embodiment of PDC 240, OHDB 248 is managed in a cyclic mode. The volume of each ODB 249 is divided into a plurality of sub-ODB. Each sub-ODB is associated with a transfer period. A transfer period, for example, may be the time interval between transferring the data of ELOG 242 to OHDB 248 and the number of sub-ODB may be a configurable number that depends on the volume of each ODB 249 and the number of records for sub-ODB that are needed. The number of records in sub-OBD may vary according to the mode of operation, i.e. “learning” or “ongoing” mode, for example.
It is conceivable that the number of records copied from ELOG to a sub-ODB may be larger than the size of the sub-ODB. In such a scenario, dropping of records may be required. It is anticipated that in order to keep the integrity of the sampled data when record dropping is required, the sample ratio of success-records copied to ODB 249s and the sample ratio of the fail-records (not having a success indication) copied into ODB 249f must be taken into consideration while preparing the predictive model. More information on the operation of PDC 240 is disclosed below in conjunction with
Each POSM 250 is associated with a web-page that is stored in one of the content servers 152 (
In some embodiments of CAS 2000 the bank of current predictive models includes a plurality of object's predictive model (OPM) 257a to 257k. Each object current predictive OPM 257 is associated with an object from the ‘a’ to k′ objects of the page and the KPI according which the model was created. This model can be replaced automatically when a model which is based on other type of KPI is ready.
A plurality of predictive factors may be retrieved by OSP 255 from the relevant entry in AST 215 and stored in PFB 253. Different methods may be used for determining which configuration of a delivered web-page maximizes the expected predicted value obtained by presenting the configuration. In one exemplary method for defining the predictive configuration, each one of the OPM 257a-k may be processed either singularly or in parallel. Per each OPM 257a-k, one or more relevant predictive factors are copied from PFB 253 into the relevant location in OPM 257 and the predictive value of the object, as well as the predictive value of the configuration (set of objects and slots), is calculated and written in a table of prediction values. After calculating the predictive value of a first object and the recommended configuration of the page in view of the first object, OSP 255 may repeat the process for the remaining one or more objects.
After calculating the predictive values and the configuration per each object, OSP 255 may scan the table of prediction values in order to determine which web-page configuration (set of slots and objects) without conflicts has the highest probability of being selected by a surfer. The identified configuration may be stored in the relevant entry of AST 215 and indication that a selection was made may be sent to MLFH 230.
An alternative embodiment of OSP may use another exemplary method, such as exhausting search, for defining the predictive configuration for the requested web-page. In such an embodiment, the table of prediction values may include an entry per each possible configuration (permutation) of the web-page. Each entry in the table conceivable may have a plurality of fields wherein each field may be associated with an optional slot. Per each cell (a junction between a row and a column in the table), the OPM 257 of the object that is associated with the slot may be fetched and calculated in view of the other configuration pairs of the web-page. The predictive value of the configuration of the web-page may be calculated as the average of the predictive value of each cell in the row such that the configuration with the highest value may be selected as the predictive configuration. More information on the operation of POSM 250 is disclosed below in conjunction with
MPM 260 manages the operation of CAS 200. It may comprise a prediction model builder (PMB) 262, a manager module 264 and a predictive model monitoring module (PMM) 266. Exemplary manager 264 may communicate with the administrator of PCD 150 (
Among other tasks, manager module 264 may define the operation mode of CAS 200. It may determine whether to work in a learning (training) mode or in ongoing mode, for example. In some embodiments the operation mode can be defined per a certain web-page. Yet in alternate embodiments, the operation mode can be defined per web-page and per KPI. Further, manager module 264 may get monitoring reports from PMM 266 and generate a decision as to whether the current predictive models are valid, need tuning or need replacing. Manager module 264 may further communicate with the administrator of PCD 150 (
On a configurable schedule, manager module 264 may scan the AST 215 and identify any inactive sections. An exemplary scanning period may be in the range of few minutes to a few hours. An inactive section is a section that the time period between the last received request (the time associated with the last entry of the section) and the scanning time is longer than a certain configurable value. Manger module 264 may sort the entries of each inactive section according to their associated web-page. Each identified entry in the inactive section may be copied into the ELOG 242 that is associated with the PDC 240 assigned to the specific web-page. After the data of the inactive entries of AST 215 has been copied, the inactive entries may be released. More information on the operation of Manager module 264 is disclosed below in conjunction with
An exemplary PMB 262 may be operable to create a plurality of predictive models wherein each predictive model may be associated with an object that may be presented in a certain web-page. In embodiments of CAS 2000 each of the predictive models may be associated with an object that may be presented in a certain web-page and a KPI that is associated to that model and that object. In some embodiments of CAS 200 or 2000, when the visit is considered while preparing a predictive model, the predictive model depends also on the number of times the relevant web-pages was requested during a certain visit.
A predictive model may include one or more predictive formulas. In order to calculate a predictive formula, an exemplary PMB 262 may use a known predictive algorithm such as, but not limited to, logistic regression, linear regression, decision tree analysis, etc. Those and others predictive algorithm are well known to a skilled person in the art and will not be further described. Further, some exemplary embodiments of PMB 262 may use stepwise methods while calculating the predictive formula. An exemplary predictive model may include one or more predictive factors wherein each set of values of the predictive factor may be associated with a coefficient in a predictive formula. Exemplary value sets may be a list of values or a range values. The coefficient may represent the effect of a predictive factor on the prediction that the relevant content object will be selected by a surfer. Exemplary predictive factors may be such as, but not limited to, the day in the week, the hour, the browser type, the number of visits to the site, the content object that is presented in another selecting button in the same web-page, the time from the last visit, a surfer's attribute, a content object attribute, etc. The different predictive factors can be processed by protocols such as IP protocol, text protocols, and time protocols. Some of the predictive factors can include set of words, such as search words, etc.
PMB 262 may use one or more sets of properties which can be used during the preparation of the predictive model. Each set of properties may be composed of a set of parameters that can be used while preparing a prediction model. An exemplary parameter may be an “aging” parameter that assigns a weight to an entry based on its age. Other parameters may define the minimum number of appearances required for a predictive factor to be considered relevant. Another parameter may define the minimum predictive score value required for a predictive model to be processed further, and so on.
The process of preparing an object's prediction model may include organizing the raw data from success and fail ODB 249 into an object's table. Each entry from the ODBs 249Ns & 249Nf is copied into a row (record) in the object's table. The rows are sorted by time, regardless of success indication. Each column in the object's table may be associated with a parameter of the record, which is stored in the row. Exemplary columns may be the representative weight of the record (the weight may reflect the relative dropping value of the ODB from which the record was copied), the result, success or failure designation, or various predictive factors that are stored in the relevant record. Exemplary predictive factors may include relevant URL keys that were embedded within the associated information that was stored in the record wherein some of the URL keys may include attributes of the requested web-page such as behavioral information values In some embodiments, the cookie may include one or more attributes of a surfer and those attributes may be associated with certain columns in the object's table. The object's table may be divided into a validation table and a training table. The training table may be referred as an analyzing table.
After organizing the object's table, a bin creating process may be initiated. A bin may be a set of values of a certain predictive factor that differs from the other sets of values, of the predictive factor, by a certain variance in prediction efficiency. The variance can be an explained variance and can represent a significant value. When the predictive factor is nominally scaled, then a bin may be a list of names. When the predictive factor is ordinarily scaled, then the set of values may be an interval of values. The explained variance threshold and the significance desired may be parameters in the set of properties. The significance can be based on difference from an average value, for example. The overall prediction efficiency of a bin may be calculated as the percentage of success compare to the total number of records in the training table that belong to the same set (the same bin), while taking into account the representative weight of the records when the bin is true and when the bin is false.
An exemplary bin may be the time interval between 8:00 am to 2:00 pm, for example. A bin table is created based on the training table by dividing each column of the training table into one or more bins (each bin represented by a column in the bin table). Each cell in the bin's table may represent a true or false value. The value may be “true” if the value in the relevant cell in the training table is in the bin's interval and false if the value in the relevant cell in the training table is not in the bin's interval.
In addition, a legend may be associated with the bin's table. The legend may define the predictive factor from which the bin was processed, the interval that the bin includes and the predictive efficiency of the bin. The process of creating a bin's table is repeated per each set of properties.
The bin's table and its legend may be loaded into a predictive model processor. The predictive model processor crunches the numbers according to a well known predictive algorithm. Exemplary predictive algorithm can be such as, but not limited to, logistic regression, linear regression, decision tree analysis, etc. Some exemplary embodiments of the present disclosure can use linear or logistic regression, with or without stepwise methods, while calculating the predictive formula. The calculated predictive model that was created based on the first set of properties is stored. PMB 262 may repeat the actions of creating a second bin's table and a second predictive model based on the second set of properties. The second predictive model may be stored as well and PMB 262 may continue the process for the remaining property sets. An exemplary PMB 262 that is installed in an exemplary CAS 2000 may repeat the actions of creating a second bin's table and a second predictive model based on the second KPI. The second predictive model may be stored as well and PMB 262 may continue the process for the remaining KPIs.
When a set of predictive models is calculated, one per each set of properties, each one of the predictive models is applied on the records of the validation table. The results of the predictions are compared with the actual recorded responses of surfers and a prediction score is granted to the predictive model. An exemplary prediction score may be the percentage of the successful predictions, when compared to the actual recorded responses, out of the total records in the validation table. The group of calculated predictive models, one per each set of properties, and or KPIs may be sorted according to the respective prediction scores. The models having a score below a predetermined minimum (one of the parameters in a set of properties) may be ignored. Then an object prediction model is calculated as a weighted average of the predictive models having prediction scores above the predetermined minimum.
In order to reduce the number of ignored prediction models, another exemplary PMB 262 may further process the comparison between the prediction results of the different models in order to adjust the associated property sets. For example, if one or more predictive models have a prediction score below the limit and use property sets in which the half-life time constant has a large value, the value may be adjusted to a smaller one, for example.
PMB 262 may repeat the above described process for each one of the possible objects in a certain web-page and then it may repeat this process for each web-page. In an alternate embodiment of the present disclosure, MPM 260 may comprise a plurality of PMB 262. Each PMB 262 may be associated with a web-page, for example. Embodiments of PMB 262 that are embedded within exemplary CAS 2000 can repeat those actions for creating predictive models per each KPI. More information on the operation of PMB 262 is disclosed below in conjunction with
An exemplary PMM 266 monitors the plurality of predictive models wherein each predictive model may be associated with an object that can be presented in a certain web-page. PMM 266 may be adapted to monitor the performance of each one of the object's prediction models that are currently used by CAS 200. Per each web-page, an exemplary PMM 266 may periodically monitor the ELOG 242 of the PDC 240 that is assigned to the current monitored web-page. The period between consecutive monitoring may be a configurable period that depends on the mode of operation of CAS 200. In “learning” mode (training mode), the configurable time period, Tmt, may range from a few minutes to tens of minutes. During “ongoing” mode, the configurable period, Tmo, may be tens of minutes to a few hours, for example.
An exemplary PMM 266 that is installed in an exemplary CAS 2000 may monitors the plurality of predictive models wherein each predictive model may be associated with an object that can be presented in a certain web-page and one type of KPIs.
Per each object that may be associated with the web-page, PMM 266 may scan the records in ELOG 242 looking for records in which a delivered web-page presents the relevant object to be selected by a surfer. Each found record may be subsequently parsed in order to determine whether the record was executed by a “common” surfer (a surfer for whom the object was presented as a result of using the object's prediction module) or a “control” surfer (a surfer for whom the object was selected randomly to be presented). The record may be further parsed in order to determine the response of a surfer, whether the object was selected by a surfer (success) or not (failure). At the end of the process, PMM 266 may calculate two probability values per each object's prediction model, i.e. the probability of success of a “common” surfer and/or the probability of success of a “control” surfer. The two probability values per each object's prediction model in the web-page may be written in a web-page comparison table. The web-page comparison table may be transferred to the manager module 264 and/or PMB 262. After processing the records stored in ELOG 242, an indication may be sent to OMM 246 of the same PDC 240 informing the OMM 246 that it may initiate the process of transferring the information from ELOG 242 to OHDB 248.
An exemplary PMM 266 that is installed in an exemplary CAS 2000 may repeat the actions of monitoring a predictive model per each KPI. In such embodiment the web-page comparison table may be associated with a web page and one of the KPIs.
At this point, PMM 266 may repeat the process for the next web-page and continue in a loop. In an alternate embodiment of the present disclosure, MPM 260 may comprise a plurality of PMM 266 wherein each PMM 266 may be assigned to a web-page. More information on the operation of PMM 266 is disclosed below in conjunction with
In addition to the above mentioned elements, exemplary embodiments of CAS 200 or CAS 2000 may comprise a non-transitory storage medium readable by a processor and storing instructions for execution by the processor for performing the capabilities of the exemplary CAS 200 or CAS 2000. The storage medium can be included as a part of a computer system of CAS 200 or CAS 2000 or sold separately.
The non-transitory storage medium can be an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system (or apparatus or device) or a propagation medium. Examples of a computer readable medium include a semiconductor or solid state memory, magnetic tape, a removable computer diskette, a random access memory (RAM), a read-only memory (ROM), a rigid magnetic disk and an optical disk. Examples of optical disks include compact disk-read only memory (CD-ROM), compact disk-read/write (CD-R/W) and DVD, etc.
Method 300 may be initialized 302 upon “power on” status or after resetting the CAS 200 and may run in a loop as long as the CAS 200 is active. During the initialization process 302 & 304, the manager 264 may be loaded with information such as information on relevant resources that stand for the internal modules of CAS 200. Accordingly, the manager 264 may allocate resources to the other modules such as PRM 220, PDC 240, POSM 250, etc. The AST 215 (
When an embodiment of method 300 is implemented in CAS 2000, the loaded information may comprise information regarding two or more KPI. This information can include one or more set of parameters per each KPI. Parameters such as value of each KPI for each object, “aging” for each KPI, per each KPI the minimum number of appearances required for a predictive factor to be considered relevant. Another parameter may define per each KPI the minimum predictive score value required for a predictive model to be processed further, information when to switch from a predictive model based on one type of KPI to a predictive model based on another KPI which is more important to the website owner, etc. Accordingly, the manager 264 may allocate resources to the other modules such as PRM 220, PDC 2400, POSM 250, etc. The AST 2150 (
After setting the internal models, the manager 264 may get 304 updated information from the content servers 152 and/or the administrator of PCD 150 (
At the end of the initiation process, MLFH 230 (
At such point, method 300 may wait 310 until the value of ST reaches a configurable value predefined to correspond to the “Learning Period”. The “Learning Period” may be in the range of few tens of minutes to a few days, depending on the volume of requests sent toward PCD 150 (
In some embodiments of CAS 2000, in which method 300 is implemented for two or more KPIs, a sub-process that includes actions 306 (
If 320 a set of object's predictive models is ready, then the set is loaded 324 into POSM 250. Other internal modules such as PRM 220, MLFH 230, OMM 246, PMB 262, OSP 255 and PMM 266 (
In some embodiments of CAS 2000, in which method 300 is implemented for two or more KPIs, each sub-process can be associated with a type of KPI. Each sub-process, which includes actions 306 (
In an alternate embodiment of the present disclosure, the monitoring mode may be implemented on a page level and not on an object level. In such an embodiment, the two counters, ST and a model's lifetime counter, are reset. The model's lifetime counter may be used for defining the overall time limit or maximum age of the predictive models. A typical, and exemplary, maximum age may be 24 hours, for example.
Then method 300 waits 330 until the value of ST reaches a configurable value predetermined to define the “Monitor Period.” The “Monitor Period” may be in the range of few tens of minutes to a few hours, depending on the volume of requests which were sent toward PCD 150 (
In some embodiments of the present disclosure, in which the monitoring mode is implemented at a page level, the probability values may be calculated at such level. In such an embodiment, a success of a page may be defined as responding to one of the presented objects. In some cases, each observation may be weighed by the percentage of the utility value of the selected object from the utility value of the page itself. In some embodiments of CAS 2000, in which method 300 is implemented for two or more KPIs, per each sub-process the selected utility value can reflects the relevant KPI associated with that sub-process.
The manager module 264 (
For CAS 2000, which handles two or more KPIs, action 332 and 334 can be executed per each KPI. Thus, when a predictive model is ready for one of the KPI, an exemplary CAS 2000 can start the ongoing period using the ready predictive model based on that KPI, while continuing the learning mode for the other KPIs. For example, due to the difference in the appearance of “clicking” versus purchasing, a predictive model based on a “click” KPI can be ready before a predictive model based on “revenue” KPI. In such a case CAS 2000 may start using the predictive model based on “click” KPI and later, when a predictive model based on “revenue” KPI is ready, CAS 2000 can switch and use the “purchasing” KPI model.
Method 300 may wait 342 until the value of ST reaches a configurable value predefined to correlate with the “Ongoing Period.” The “Ongoing Period” may be in the range of a few tens of minutes to a few days, depending on the volume of requests which were sent toward PCD 150 (
When 342 the ST reaches the value of the “Ongoing Period,” then PMM 266 (
At action 354, the value of the model's lifetime counter is checked and a determination is made 360 as to whether the value is below a configurable value Tmax. Exemplary Tmax may be a few hours to a few days, for example. If 360 the value is below Tmax, which would indicate that the sets of object's predictive models are still valid, then the ST counter may be reset 364 and method 300 returned to action 342. If the value of the lifetime counter is above Tmax, then method 300 may proceed to action 362 and returned to action 306 (
During the initialization process 402, PRM 220 may be loaded with information that is relevant for communicating with other internal modules of CAS 200, modules such as the manager 264 and AST 215 (
After the initiation stage, method 400 may start processing surfer's requests that are transferred via HTTP proxy 210 (
At action 412 the first request of a new active surfer is handled. An ID is allocated to the new surfer. The ID can be allocated randomly from a large number of alternatives. The ID may be a 32 bit, 64 or 128 bit number, for example. In case that the current mode is a learning mode, then the surfer is defined as a random surfer. In case that the current mode of operation is monitoring mode or ongoing mode. The type of the new surfer as a personalized surfer or as a random surfer can be defined based on the allocated ID and a percentage of surfers per each type. In the monitoring period, the percentage of the random surfers is larger than the percentage of the random surfer that is associated with the ongoing mode. Thus, the portion of the random surfers in a monitoring period is larger than the control portion during the ongoing mode. The value of percentages is a configurable value depending on the traffic via CAS 200.
A new section in AST 215 (
In some embodiments that calculate predictive model per visit, the receiving time of the first request can indicate the beginning of the visit. In some embodiments this time can be stored in a dedicated filed of the new section that is associated with the new requester's ID. In some exemplary embodiments of the present disclosure, the URL may include one or more fields that reflect attributes of the requested web-page. Exemplary attributes may be the topic of the page (vacation, news, sports, etc.), classification of the advertised goods, (expensive, low cost, etc.).
The decompressed cookie, if it exists, is updated 412 or a new cookie may be written. The cookie may include behavioral information associated with the surfer such as the date and time of the surfer's last visit to the web-page, the frequency of visits by the surfer to the web-page, etc.
Other behavioral information may include one or more counters wherein each counter may count the number of events of a certain type. Exemplary counters may count the number of the active surfer's visits to the relevant website, the number of requests for a certain web-page from the website, the number of times a certain offer (content object) was selected, etc. Updating the counters may be time dependent Such that the value of the counter may decrease over time, as previously described. In some embodiments, the cookie may include labels for a surfer's attributes. Exemplary attributes may be a surfer's age, gender, hobbies, etc. The updated information of the cookie is stored in the appropriate fields of the entry in AST 215.
In exemplary embodiments that create predictive model based on the entire visit of surfer an additional counter can be added for counting the number of requesting the relevant webpage in the relevant visit. This counter can be reset each time a new visit is initiated by the relevant surfer to that website.
After setting the fields in the section of AST 215, the method 400 proceeds to action 456 in which the requester ID is written in the cookie of the request and the request is transferred toward its destination (one of the servers 152 in
In an alternate exemplary embodiment of the present disclosure, defining a surfer as a random surfer may be executed after receiving a surfer's response. In such an embodiment, a surfer that responded positively may be defined as a random surfer. One of the reasons for using this sorting method is to increase the size of the positive samples, since the positive samples are less populated than non-responding surfers.
In embodiments of CAS 2000 (
Returning now to action 432, if a section that is associated with the requester's ID exists, then a surfer's section in AST 215 (
After writing the information in the new entry, the URL that is associated with the request is parsed 442 and the requester's section in AST 215 is searched 442 for an entry that is associated with a previously delivered page (PDP) that points to the current requested object in order to determine whether the current request originated from a web page that was recently sent via CAS 200. Presumably, such a web-page included an object and slot combination which reflects the URL that is associated with the current received request.
If 450 such an entry is found, then a success indication is written 452 in the result field that is associated with the object and slot combination of the URL associated with the current request. At such point, method 400 proceeds to action 456. If 450 a PDP is not found, method 400 proceeds to action 456.
There are some exemplary embodiments in which two or more KPI are used or in which the entire visit is taking into consideration. Such exemplary embodiment may record the entire visit of a user, based on the user ID and supplement ID that reflects the visit. The record can include information on the surfer's requests during the visit, the responses sent per each requests, the result of the visit (one or more purchases done by the surfer during the visit, the revenue per each purchases and the total revenue, for example). At the end of the visit the recorded information is parsed and added to the relevant entries and KPIs in the surfer's section in the AST 215.
During the initialization process 502 & 504, OMM 246 may be loaded with information that is relevant for communicating with other internal modules of CAS 200 such as the manager 264, PMM 266, ELOG 242, (
After the initiation process, a determination is made 506 as to whether the current operating mode is a learning mode (training mode). If not a learning mode, method 500 may wait 507 for an indication from PMM 266, in which PMM 266 may inform OMM 246 that the data, which is stored in the current ELOG 242, was inspected and that OMM 246 may start transferring the data from the ELOG 242 toward OHDB 248 (
At action 509 a new ELOG 242 is allocated 509 for storing the future events that are related to the relevant web-page and method 500 begins processing and transferring the stored data from the old ELOG 242 into OHDB 248. A loop from action 510 to action 520 may be initiated and each cycle in the loop may be associated with an optional object that may be presented over the relevant web-page.
At action 512, the oldest sub-ODB in each one of the ODB couples associated with the current object, object “a” for example which comprises the success ODB 249as and the failure ODB 249af, are released and a new sub-ODB is allocated for storing data that is transferred from the old ELOG 242 that is related to the current object. The old ELOG 242 (
The portion of the entries from each group that may be stored in the appropriate new sub-ODB is calculated 514 by dividing the number of entries in each one of the appropriate new sub-ODB (success or failure) by the total number of entries in each group (success or failure, respectively). If at least one of the groups has a portion value smaller than one, meaning that some of the entries of the group will be dropped in lieu of being stored in the new sub-ODB, then the smallest portion value may be selected for determining the number of entries from each group (success or failure) that will be stored in the appropriate sub-ODB (success or failure, respectively).
The total number of entries in each group is multiplied by the smallest portion value for determining the numbers of entries (NE) from each group that will be stored in the appropriate new sub-ODB. NE entries from the top (the newest entries) of each group are copied 516 to the appropriate sub-ODB. If for both groups the portion values are greater than “one,” then all the entries of each group (success or failure) may be copied to the appropriate new sub-ODB (success or failure, respectively). The portion value of the deletion per each group (success and failure) may be recorded in association with the success ODB 249as-ns and the failure ODB 249af-nf (respectively). In some exemplary embodiments of the present disclosure, old entries in the success ODB 249as-ns and the failure ODB 249af-nf may be released only when there is no free space available for storage of new observations.
After copying 516 the entries to the new sub-ODB, a determination is made 520 as to whether additional objects may be presented in the relevant web-page. If yes, method 500 may start a new cycle in the loop for the next object and return to action 510. If 520 there are no more objects, then the old ELOG 242 is released 522, timer OT is reset, and method 500 returns to action 506.
In some embodiments of CAS 2000, in which method 500 is implemented for two or more KPIs, each sub-process can be associated with a type of KPI. Each sub-process can include actions 506 to 522. Such an embodiment of CAS 2000 may have a separate ELOG 242 (
In order to eliminate redundancy of information in the following paragraphs, actions of method 600 that implemented in CAS 200 or CAS 2000, which are similar, are disclosed once, in relation with CAS 200. Base on reading the description of method 600 for CAS 200, an ordinary skilled person in the art is expected to understand how those similar actions are operating in relation with CAS 2000. In some actions method 600 in which CAS 2000 may operate in a different way or may need additional information, this information will be added.
After the initiation stage, method 600 may start processing received web-pages from the content servers 152 (
If 610 a surfer is designated as a random surfer, then at action 612, based on the information written in the entry in the AST associated with the web-page ID, a list of optional objects and a list of slots in the web-page are fetched. Per each slot, an optional object is selected randomly and method 600 proceeds to action 622. In an alternate embodiment, in which a surfer type (random surfer) is defined per slot, then the random selection is done only for that slot.
If 610 the requester of the web-page is not designated as a random surfer, then a POSM 257 is selected 614 based on the page ID. The entry number in AST 215 that is associated with the request for the given web-page is transferred to a queue of the selected POSM 250. Method 600 may wait 620 until a decision from the relevant POSM 250 is received that identifies the web-page configuration that has the highest probability of prompting an active surfer to respond through the selection of one of the optional objects. The identified configuration may include a list of slot and object combinations that have the highest associated probabilities to highest utility to the web-site owner. Then method 600 proceeds to action 622.
At action 622, either the identified configuration of the web-page or the randomly selected configuration of the web-page, having a set of object and slot combinations, is written 622 to the appropriate fields in the relevant entry of AST 215. The HTML file is modified to include the selected one or more objects, each object being associated with an appropriate slot. The updated cookie, which was prepared by PRM 220, manipulated by the relevant content server 152 (
Yet in an alternate embodiment of CAS 200 or 2000, the selected configuration is sent by the API to the content server 152 (
In an alternate embodiment, the set of pairs in the configuration may influence the design of a web-page. In such an embodiment, a pair may include a design feature in lieu of a slot and the value of the design feature in lieu of the object. Exemplary design features may be background color, font size, font type, etc. Exemplary corresponding values may be red, 12 points, Times New Roman, etc.
In order to eliminate redundancy of information in the following paragraphs, actions of method 700 that implemented in CAS 200 or CAS 2000, which are similar, are disclosed once, in relation with CAS 200. Base on reading the description of method 700 for CAS 200, an ordinary skilled person in the art is expected to understand how those similar actions are operating in relation with CAS 2000. In some actions method 700 in which CAS 2000 may operate in a different way or may need additional information, this information will be added.
After initiation, method 700 may fetch 704 the entry from AST 215 that is associated with the request for this delivered web-page. The entry may be parsed and a plurality of predictive factors can be retrieved from the associated information, behavioral information, and grouping information (if such exists). Exemplary predictive factors that may be retrieved from the associated information stored at the entry may include the receipt time of the request, the URL keys associated with the request, etc. Exemplary predictive factors that may be retrieved from the behavioral information may include the elapsed time from the last visit of the requester in PCD 150 (
The optional configurations of the requested web-page are calculated 704. A configuration may define a set of the slots, in which an optional object may be presented, and a selected optional object that is presented in each slot (Slot/Optional-object). Usually, an exemplary web-page has a certain amount of slots and a certain amount of optional-objects such that the number of possible configurations may be defined. The optional configurations may include all possible combinations of slots and objects, for example. A table of prediction values may be allocated 704 for storing the calculated prediction value for each optional configuration.
An exemplary table of prediction values may include a plurality of entries. Each entry may be associated with an optional configuration of the delivered web-page. Further, each entry may include a plurality of fields and each field may be associated with a specific slot number with an additional field defining the predicted value of each configuration. Each cell (a junction of a configuration and a slot number) in the table may include the calculated prediction value of the object that is assigned to the slot in that configuration, wherein the calculation of the predictive value is done in view of the rest of the slots and optional-objects in the configuration. The prediction value of the configuration may be the sum of the predictive values of each one of the slot fields in this entry.
In some exemplary embodiments of the present disclosure, the number of optional configurations may be calculated once, during the initialization of POSM 250 or upon receiving information on changes in the associated web-page. In such case, an exemplary table of prediction values may be defined once and allocated again and again each time the relevant web-page is delivered.
After allocating the table of prediction values, a loop between actions 710 to 740 may be initiated to process all the possible configurations of the web-page written in the allocated table. For each configuration, an internal loop is initiated between actions 720 to 730. The internal loop may be executed per each slot of the configuration. An OPM 257a-n (
Then the predictive value of the object that is associated with the first slot of the first configuration is calculated 724 by executing the model. Next, the calculated predictive value is written in the table's cell at the junction of the first configuration and the first slot and a decision is made 730 as to whether there are additional slots in the configuration. If so, method 700 returns to action 720 and starts the loop for the next slot (a second slot, etc.) in the first configuration. The loop may continue to the third slot and so on until there 730 are no more slots in the configuration. Then, the predictive value of the configuration is calculated 732 by summing the predictive value of each slot, for example. In other embodiments the predictive value of the configuration is calculated 732 by averaging the predictive value of each slot. etc.
At action 740, a determination is made as to whether there are additional optional configurations in the table. If yes, method 700 returns to action 710 and starts the loop for the next configuration (a second configuration, etc.) in the table. The loop may continue for the third configuration and so on until there 740 are no more configurations to be processed.
After calculating 740 the predictive value of all the possible configurations of the web-page, an optimal configuration is selected 742. Notably, as previously disclosed, the prediction value also reflects the benefit that the owner of CAS 200 will get when a surfer responds to the presented web-page via the selection of one of the optional objects. The configuration with the highest predictive value may be defined as the preferred configuration. This preferred configuration, and its predictive value, may be stored 744 in the entry of AST 215 (
In an exemplary embodiment of CAS 2000 in which two or more KPIs are used. Method 700 can be implemented by using a ready predictive model that is associated to a KPI with the highest utility to the owner of the web-site. Such embodiment may start using the predictive model in which the KPI reflects clicking on a certain object. Later it may switch to a predictive model that is associated with purchasing KPI, for example.
Method 800 may be initialized 802 by the manager module 264 (
After initiation, a timer Tm is reset 804 and the current mode of operation, which is defined by the manager module 264 (
When timer Tm reaches the value of Tmm or Tmo, depending on the operating mode, a web-page comparison table is allocated 820. The web-page comparison table may have two entries wherein one may be assigned to the average utility of the random surfer while the other may be assigned to the average utility of the personalized surfer. An exemplary method 800 which is implemented in a CAS 2000, for example, can have two or more web-page comparison tables. Each web-page comparison table can be associated with a KPI.
At action 836, the ELOG is scanned and the average utilities of the random surfer and the personalized surfer are calculated. The results may be written in the web-page comparison table at the appropriate entry that is assigned to the relevant surfer.
The web-page comparison table is transferred 842 to the manager module 264 (
Method 900 may be initiated 902 by the manager module 264 (
When the initial action 904 is completed, an external loop between actions 910 and 960 (
In exemplary embodiments of process 900 that are implemented by CAS 2000 the external loop between actions 910 and 960 (
The raw data from the success and failure ODB 249Ns & 249Nf are organized 914 into an object's table. Each entry from the ODBs 249Ns & 249Nf is copied into a line in the table. The lines are sorted by time, independent of any success indication. The newest record may be stored at the top of the object's table while the oldest record may be stored at the bottom of the object's table, but it should be understood that choice of storage organization should not limit the scope of the disclosure.
Each record in the table, i.e. line, has a plurality of columns (fields). Each column may be associated with a factor of the record, which is stored in the record. Exemplary columns may be designated to represent the weight of the record (the weight may reflect the representative weight of the ODB from which the record was copied), the result, success or failure (not success), relevant URL keys that were embedded within the associated information that was stored in the record, attributes indication on the web-page and/or the object, etc. Additional columns may reflect the values of behavioral information, which is stored in the record. Other columns may represent information such as counters and timers that were stored in the cookie associated with the request, indications of attributes associated with a surfer that were stored in the cookie, etc. A column can reflect the utility of the success to the website owner. The utility can be one of the KPIs. In some embodiments in which the entire visit of a surfer is taken into consideration, then a column can be associated to the number of times the surfer, that is associated with the record, requested the relevant web page during the same visit.
The associated information may be the receiving time of the request that is written in the record, the URL that was associated with the request, information regarding the configuration of the web-page from which the request was selected, etc. Each cell at the junction of a column and a row may store the URL key's value if it exists, missing key indication if the record does not include the key, or a missing value indication if the record includes the key but no value, for example. Each cell in the junction of a behavioral information factor and a line (record) may store the value of a relevant counter, timer, etc. which are associated with the column.
The object's table may be searched 914 for irrelevant keys (few instances), for example. Irrelevant keys may be defined as keys that have a small number of records (lines) in which the key has a value. The minimum number of records may be a configurable value in the range of a few tens to a few thousands, depending on the volume of data stored in the ODB 249, for example. The minimum number of records may be one of the properties that are stored in a set of properties, for example. A column of irrelevant keys may be removed from the object's table. The object's table may be divided into two tables: a validation table including the newer records and a training table with the older records. Usually, the training table includes more records than the validation table. The validation table may be used later for determining the quality of the predictive models or the score of the model.
The training table is further processed 916 in order to calculate the predictive model of the object. The type of the columns in the training table may be defined. Defining the type may be executed by observing the key in view of a plurality of syntax protocols and determining whether the value of the key complies with one or more of those protocols.
Next, the type is defined 916 based on the protocol. Exemplary protocols may be Internet protocols (IP), text protocols, time protocols, etc. If the value of the key does not comply with any of the protocols, it may be defined as “other.” Exemplary types may include IP, time, a list of searching words, text, numbers, currency, other, etc. Types of columns that are associated with behavioral information are known and may include counters and timers, for example. After defining the type of the keys, the type of the value of each key may be defined. Exemplary types of values may be scale, ordinal, nominal, cyclic, etc.
In some exemplary embodiments, method 900 may further process 916 the training table in order to identify columns that may be divided into two or more columns. For example, a time column may be divided into multiple columns representing days, hours, and minutes. An IP address may be divided into four columns, etc. In other embodiments, an IP address can be stored in a single column using 12 digits, padding with leading zeros and removing the points. Each sub-column (a portion of a predictive factor) may be referred to as a predictive variable or a predictive key.
When the raw data is organized in the training table, method 900 starts converting 916 key values that are not ordinal into ordinal values. A value of a text column (a certain string of text) may be converted into an ordinal number that reflects its frequency (number of appearances along the column). A nominal number or a cyclic number may be converted into an ordinal number that reflects the influence of the nominal value on the result of a record (success or failure). In other embodiment a text or nominal values can be represented in an ordinal way by sorting the predictive power of each value. The predictive power can be defined as the ability to generate testable predictions.
At this point, the training table is ready to be further processed such that each column (a predictive variable/factor/key) is converted into one or more bins and the training table converted overall into a bins table. In addition, a legend may be associated with the bin's table. The legend may define the predictive factor from which the bin was processed, the interval that the bin includes and the predictive efficiency of the bin. The bins table is further processed for determining a predictive model per each set of properties. A middle loop from action 920 to 950 (
An exemplary parameter may be relative aging weight. Other parameters may define the minimum number of appearances of a certain key, a number below which a key may be considered as irrelevant. Another parameter may define the minimal value of a predictive score that a predictive model may get in order to be used. Yet another parameter may define the half-life-time constant of a record, etc. In an exemplary embodiment of method 900, which is implemented in CAS 2000, each KPI may have its own set of properties. A KPI that reflects clicking rate may have different parameters than KPI that reflects purchasing. For example the minimum number of appearances for clicking rate can be higher than the minimum number of appearance per purchasing, etc.
The value of M counter is incremented 922 by one indicating the ID number of the set of properties that is associated with the current cycle of the loop. The appropriate set of properties is fetched and parsed and, according to its properties, an equivalent weight per each line may be defined. The equivalent weight may reflect the representative weight which is due to the result of the record's (line) success or failure, age, the benefit that is created to the owner of CAS 200 (
In some embodiments, the binning process is carried out with respect to a target KPI. The target KPI can be Boolean (0 or 1), discrete, continues or mixed as well. The binning process calculates the average KPI within the bin and compares it to the average KPI outside the bin, while calculating the significance of the difference. This approach may also be used to split a column into two bins, generating the highest explained variance, and further splitting of each bin. Each split of a bin may have some constraints like certain allowed percentiles split and minimal size, etc.
Each cycle in the internal loop is associated with a column, i.e. a predictive variable. The range of the ordinal values written in the column is divided 932 into one or more sub-intervals (bins) according to the ability to predict the success. Dividing the range of the values of the column into bins may be executed in several methods. One exemplary method may divide the interval into a configurable number of equal intervals (units). The number of units (the resolution) may be a few units to a few tens of units, for example. The resolution may be one of the properties that are included in a set of properties, for example. A predictive score per each interval unit, along the interval of the values of the current column, may be calculated by dividing the weighted number of success records by the total weighted number of records, which have an ordinal value of the predictive variable (column), in the current interval unit.
The one or more bins may be created 932 by grouping one or more adjacent interval units into one bin, wherein the variance between the rates of success of the adjacent interval units is below a certain value, i.e. a variance threshold. The variance threshold may have a configurable value and may be one of the properties that are stored in a set of properties, for example. After creating one or more bins for each predictive variable (column), a rate of success is calculated for the bin. The rate of success is calculated as the weighted number of success records divided by the weighted total number of records. Information on the bins (information on its predictive variable, the bin's interval, etc.) along with associated prediction scores is stored in a bins legend. Then, a determination is made 940 as to whether the training table includes more columns (predictive variables). If 940 yes, method 900 returns to action 930 and starts a new cycle in the internal loop for dividing the next predictive variable (a next column, in the training table) into one or more bins. In embodiments that target a non Boolean KPI, the average KPI can replace the rate of success.
If 940 there are no more columns, then a bins table is created 942. The bins table may have a column per each bin and the same number of lines as in the training table. In each cell, at the junction of a line (record) and a column (bin), a binary value, true or false, may be written depending on the value of the relevant predictive variable. The value of the relevant predictive variable is in the range of values that is associated with bin.
In another exemplary embodiment of method 900, action 942 may include a validation process. An exemplary validation process may repeat the loop from action 930 to 940 on the records that are stored in the validation table. If the bins that were created by processing the validation table are similar to the bins that were created by processing the training table, then the bins may be considered as valid and method 900 may proceed to action 944. If not, one or more properties in the current set of properties may be slightly modified. For example, the resolution of the range of a certain column may be reduced. Then, the loop from action 930 to 940 may be repeated twice with the modified set of parameters. During the first repetition a second set of bins is calculated using the records stored in the training table. During the second repetition, the second set is validated by using the validation table. The validation process may have one or more cycles. In some exemplary embodiments the bins validation table may have other records than the validation table that is used for validating the predictive models.
Yet another exemplary embodiment of method 900 may respond to non valid sets of bins by marking the set of properties as a problematic one and may jump to action 950, skipping the stage of calculating a predictive model according to the problematic set of parameters. In yet another embodiment, a set of bins may be ignored entirely.
The bins table and the bins legend is transferred 944 toward a predictive model engine. The predictive model engine may implement a statistical algorithm over the bins table in order to calculate the predictive model. The statistical algorithm can be such as, but not limited to, logistic regression, linear regression, decision tree analysis, etc. The calculated predictive model that was created is stored as a prediction model (N;M). The N stands for the optional-object for which the model was calculated and the M stands for the set of properties that was used for calculating the model.
At action 950, a determination is made as to whether there are more sets of properties. If yes, method 900 returns to action 920 (
An exemplary predictive model (N:M) may include one or more constants, one or more predictive variables that are derived from associated information, each having an associated coefficient, one or more predictive variables that are derived from behavioral information, each having an associated coefficient, one or more web-page configuration variables that reflect objects in the other slots, each having an associated coefficient.
In some embodiments, additional predictive variables may be used. For example, one or more predictive variables that are derived from web-page attributes, surfer's attributes, or object attributes may be included. Exemplary constants may be a result of the regression process (an arithmetic constant), for example. Another constant may reflect the benefit of selecting the object M. Exemplary predictive variables that are derived from associated information may be the day, the browser type, etc. An exemplary predictive variable that is derived from the behavioral information stored in the cookie may be the number of visits to a certain page, etc. An exemplary web-page configuration variable may be object X in slot Y, for example.
An exemplary calculation of a Boolean KPI (true, fails) predictive value of an object may be executed based on the following formula:
The relevant coefficients are the coefficients that are associated with variables that are true for a received request.
An exemplary method 900 may execute 952 the M models of the optional-object N on the records that are stored in the validation table. The M models may be executed on each record in the validation table that includes the object N in a web-page configuration that was sent as a response to the request that initiated the record. After performing the M models on the validation table, a predictive fitness score is calculated per each model. In an alternate embodiment of the present disclosure, the records that are stored in the validation table plus the records that are stored in the training table may be used in action 952.
PMB 262 may select 952 a group of calculated predictive models out of the M models that have the best scores. Then, the object N predictive model may be calculated as a representative model of the group of the best models. Calculating the representative model may be executed by calculating an average value per each coefficient. The average may be weighted by the predictive score of the selected best models, for example. The representative model may be stored as “object N ready to be used predictive model”. A bin that does not exist in a model is considered to have a coefficient of zero. In an embodiment, the equivalent model can be generated by merging the M models using all the data, the training and the validation. The equivalent model will be created based on the previously calculated fitness scores. In this disclosure, the terms “equivalent model” and “representative model” can be used interchangeably.
At action 960, a decision is made whether there are additional optional-objects that may be associated with the relevant web-page. If yes, method 900 returns to action 910 and start the external loop for calculating a prediction model for a next object (M+1). If there are no more objects, Manager module 264 (
Usually each KPI has a different distribution, therefore predictive models of the same page that are based on different KPI may be ready at a different time. It is expected that the predictive model, which is based on clicking as a KPI, will be valid first because the clicking occurs more frequently than purchasing, for example. Consequently the manager model 264 (
It is to be understood that the above description is intended to be illustrative, and not restrictive. The above-described apparatus, systems, and methods may be varied in many ways, including, changing the order of steps, and the exact implementation used. The described embodiments include different features, not all of which are required in all embodiments of the present disclosure. Moreover, some embodiments of the present disclosure use only some of the features or possible combinations of the features. Different combinations of features noted in the described embodiments will occur to a person skilled in the art. Furthermore, some embodiments of the present disclosure may be implemented by combination of features and elements that have been described in association to different exemplary embodiments along the discloser. The scope of the invention is limited only by the following claims and equivalents thereof.
Claims
1. (canceled)
2. A computer-implemented method, comprising:
- generating, by a computing device, a plurality of predictive models for an application, wherein the application is associated with a configuration defining one or more objects for presentation in a window of the application, wherein an object is associated with a predictive model including one or more predictive factors, wherein a predictive factor of a predictive model is used to calculate a predictive value for the predictive model, and wherein the predictive value predicts a property of an object associated with the predictive model;
- receiving a request for content associated with the application, wherein the request includes predictive information;
- determining the predictive information associated with the request;
- determining values of predictive factors for the plurality of predictive models, wherein the values are determined using the predictive information;
- determining multiple configurations of the application, wherein each configuration includes one or more different objects for presentation with the content;
- determining predictive values for the multiple configurations of the application using the determined values of the predictive factors, wherein a predictive value of a configuration includes a combination of predictive values of predictive models associated with objects making up the configuration;
- selecting a configuration from the multiple configurations with a highest predictive value; and
- providing the selected configuration for presentation by the application.
3. The method of claim 2, wherein determining the multiple configurations of the application includes determining one or more objects to present in each configuration.
4. The method of claim 2, wherein the predictive information includes behavioral information and information associated with one or more responses to one or more objects.
5. The method of claim 4, wherein the behavioral information includes a time of a previous request associated with content of the application, a time that the content was last requested, a number of visits to the application, a number of times that an object has been presented, and a number of times that the object has been selected.
6. The method of claim 4, wherein the information associated with the one or more responses to the one or more content objects includes a receipt time of a response, a uniform resource locator key associated with the response, an internet protocol address, or a type of application.
7. The method of claim 2, wherein the application includes a mobile website.
8. The method of claim 2, wherein the application includes a mobile application.
9. A system, comprising:
- a processor;
- a non-transitory computer-readable storage medium containing instructions which when executed on the processor, cause the processor to perform operations including:
- generating a plurality of predictive models for an application, wherein the application is associated with a configuration defining one or more objects for presentation in a window of the application, wherein an object is associated with a predictive model including one or more predictive factors, wherein a predictive factor of a predictive model is used to calculate a predictive value for the predictive model, and wherein the predictive value predicts a property of an object associated with the predictive model;
- receiving a request for content associated with the application, wherein the request includes predictive information;
- determining the predictive information associated with the request;
- determining values of predictive factors for the plurality of predictive models, wherein the values are determined using the predictive information;
- determining multiple configurations of the application, wherein each configuration includes one or more different objects for presentation with the content;
- determining predictive values for the multiple configurations of the application using the determined values of the predictive factors, wherein a predictive value of a configuration includes a combination of predictive values of predictive models associated with objects making up the configuration;
- selecting a configuration from the multiple configurations with a highest predictive value; and
- providing the selected configuration for presentation by the application.
10. The system of claim 9, wherein determining the multiple configurations of the application includes determining one or more objects to present in each configuration.
11. The system of claim 9, wherein the predictive information includes behavioral information and information associated with one or more responses to one or more objects.
12. The system of claim 11, wherein the behavioral information includes a time of a previous request associated with content of the application, a time that the content was last requested, a number of visits to the application, a number of times that an object has been presented, and a number of times that the object has been selected.
13. The system of claim 11, wherein the information associated with the one or more responses to the one or more content objects includes a receipt time of a response, a uniform resource locator key associated with the response, an internet protocol address, or a type of application.
14. The system of claim 9, wherein the application includes a mobile website.
15. The system of claim 9, wherein the application includes a mobile application.
16. A computer-program product tangibly embodied in a non-transitory machine-readable storage medium, including instructions configured to cause a data processing apparatus to:
- generate a plurality of predictive models for an application, wherein the application is associated with a configuration defining one or more objects for presentation in a window of the application, wherein an object is associated with a predictive model including one or more predictive factors, wherein a predictive factor of a predictive model is used to calculate a predictive value for the predictive model, and wherein the predictive value predicts a property of an object associated with the predictive model;
- receive a request for content associated with the application, wherein the request includes predictive information;
- determine the predictive information associated with the request;
- determine values of predictive factors for the plurality of predictive models, wherein the values are determined using the predictive information;
- determine multiple configurations of the application, wherein each configuration includes one or more different objects for presentation with the content;
- determine predictive values for the multiple configurations of the application using the determined values of the predictive factors, wherein a predictive value of a configuration includes a combination of predictive values of predictive models associated with objects making up the configuration;
- select a configuration from the multiple configurations with a highest predictive value; and
- provide the selected configuration for presentation by the application.
17. The computer-program product of claim 16, wherein the predictive information includes behavioral information and information associated with one or more responses to one or more objects.
18. The computer-program product of claim 17, wherein the behavioral information includes a time of a previous request associated with content of the application, a time that the content was last requested, a number of visits to the application, a number of times that an object has been presented, and a number of times that the object has been selected.
19. The computer-program product of claim 17, wherein the information associated with the one or more responses to the one or more content objects includes a receipt time of a response, a uniform resource locator key associated with the response, an internet protocol address, or a type of application.
20. The computer-program product of claim 16, wherein the application includes a mobile website.
21. The computer-program product of claim 16, wherein the application includes a mobile application.
Type: Application
Filed: Jun 14, 2016
Publication Date: Jan 12, 2017
Patent Grant number: 11263548
Applicant: LivePerson, Inc. (New York, NY)
Inventors: Shlomo LAHAV (Ramat-Gan), Ofer RON (Givatayim)
Application Number: 15/182,310
